JP2003515942A - Power-assisted automatic monitoring and classifier fabrication tool for semiconductor defects - Google Patents

Abstract

(57) SUMMARY A method and system that optionally allows a user to view image defects organized by natural grouping based on image features. Natural grouping makes it easier for the user to organize some or all of the images into classes in a training set of images. A feature vector is extracted from each image in the training set and stored with its user-specified class for use by the automatic classifier software module. The automatic classifier uses the stored feature vectors and classes to automatically classify images that are not in the training set. If the automatically classified image does not match the image manually classified by the user, the user modifies the training set until better results are obtained from the automatic classifier. The system can provide feedback to an inspection system designed to assist in setting up and fine-tuning the inspection system.

Description

Detailed Description of the Invention

Related Application The present invention was filed on Nov. 29, 1999, Bakker, Baner.
Jee and Smith et al., "Power Assisted Autom"
atic Supervised Classifier Creation
35 US Pat. No. 60 / 167,955 entitled “Tool for Semiconductor Defects”. S. C. § 119
Claim priority under (e).

The following applications are related to the present application and are incorporated herein by reference: 1. US Application No. 08 / 958,288 to Hardikar et al., Filed October 27, 1997. 2. US Application No. 08 / 958,780 to Hardikar et al., Filed October 27, 1997.

The following US patents are related to the present application and are incorporated herein by reference: 1. U.S. Pat. No. 5,226,118 issued July 6, 1993 and issued to Baker et al.

BACKGROUND OF THE INVENTION Field of the Invention The present invention relates generally to software programs, and more generally to software controlled inspection systems for optical, e-beam, or other types of semiconductor wafers.

2. Description of the Background Art As new materials, methods and processes are introduced into semiconductor manufacturing, new defects are emerging in the manufacturing process that can significantly impact yield. These changes employ new technologies that semiconductor manufacturers can use to detect and classify defects that limit their yields faster, more accurately and consistently in order to pack their manufacturing process and shorten the yield learning curve. Is seeking. At the same time, the shrinking product life cycle and reduced time-to-market requirements help manufacturing plants speed up their product ramp-up for new products to meet their profitability goals. Have to And this ensures that the need for faster automatic defect classification (ADC) setup allows the manufacturing plant to gain the ADC's benefits without slowing down the ramp process.

Existing optical inspection systems with automatic defect classification set up an automatic classification system in which a human visually inspects a semiconductor wafer looking for suspected defects and classifying the type and cause of the defects. Wants to do. To make this classification, humans must segment hundreds of images presented in any order. This process takes hours and adds to the cost of manufacturing. Moreover, even skilled human operators are somewhat slow and error prone.

SUMMARY OF THE INVENTION The described embodiments of the invention receive a defective image and assist the user in classifying the type of defect presented. The graphical user interface allows a human user to manually classify defective images through the auto-monitoring classifier software, and further allows the user to manually classify his or her own manual classification. Allows to contrast with the classification determined by.
In order to fabricate an automated monitoring classifier for semiconductor defect classification, a human user has to perform two manual tasks:

(I) Creating a good classification scheme (which images a user puts into which classification). (Ii) Creating a good training set example for an auto-supervised classifier with this classification scheme.

The described embodiments of the invention provide tools that assist human users in achieving both of these goals at the time of recording. The embodiments described herein include the following four main components that interact seamlessly with one another:

(I) Image Gallery: This is a graphical interface that displays a list of images in a systematic fashion. (Ii) Dynamic self-monitoring classifier: Given a list of defects, allows the user to manually classify or train any set of defects. The remaining defects are dynamically classified accordingly and the overall performance of the resulting classifier is calculated.

(Iii) Dynamic Classifier Control and Performance Tool: This allows the user to visualize the classifier parameters for further optimization. (Ii
), The resulting performance will be immediately visible.

(Iv) Unsupervised automatic classifier (natural grouping): This is a method that allows the user to visualize the layout and structure of the feature space in terms of defective images. Grouping helps the user to create both good classification schemes and good training set examples.

Detailed Description of the Preferred Embodiments The described embodiments of the present invention assist a user in classifying defective images input to the system from external sources. For example, the image is a scanning electron microscope (S
EM), defect inspection system such as KLA-Tencor 2132 inspection tool,
Or it can be received from any similar device used to generate an image and detect errors in it. In particular embodiments, images may come from more than one type of input source. For example, images can be received from both SEM and 2132 inspection tools. Different types of images received from different sources help in defect classification because they provide different types of images that produce more information about the defects to be classified. is there. The image can also be the location of points on the wafer.

FIG. 1A is a schematic block diagram of a semiconductor optics, e-beam, or other type of inspection system. As mentioned above, the defective image is preferably received from an external source such as an optical, e-beam, or other type of inspection system 102, SEM or 2132 inspection tool. Classification system from defective images 10
It can be seen that a defect occurred in No. 4, but the defect regarding the type or cause is not classified. After the defects have been classified, the defects may be KLA-Tenco, as described below.
to an analyzer 106, such as the Klarity system of R.r., an example of which is referenced above in US patent application Ser. Nos. 08 / 958,288 and 08 / 958,780.
No.

FIG. 1 (b) is a schematic block diagram of a semiconductor optics, e-beam, or other type of inspection system that uses the present invention for inspection setup. As mentioned above, the defective image may be detected by an optical, e-beam, or other type of inspection system 1.
02, SEM or 2132 inspection tools are preferred. These images may be associated with etching, photolithography, deposition, CMP, or some other manufacturing process. Classification system from defective images 10
It can be seen that a defect occurred in No. 4, but the defect is not classified with respect to type and cause. After being categorized, the defects are KLA-Tencor's KLAri as described below.
The output is sent to an analyzer 106, such as a ty system, an example of which is described in US patent application Ser. Nos. 08 / 958,288 and 08 / 958,780 referenced above. The output of analyzer 106 is used as feedback to fine tune the inspection system.

The inspection system can be fine-tuned to increase or decrease the sensitivity of the system, for example, depending on the number of errors found and the desired accuracy. For example, if the system finds too many errors, or errors that are not related to a particular manufacturing process, you can fine-tune the system to reduce its sensitivity and detect fewer errors. . As another example, if not enough errors have been detected, or if a particular type of error has not been detected, to detect data that leads to the detection of more or desired types of errors. , The inspection system can be adjusted for even greater sensitivity. In certain embodiments, increasing sensitivity has been observed to increase error detection in a geometrical manner. In such a system, feedback to the ADC may be used to control inspection parameters including, but not limited to, illumination, sensitivity or sensing (optical, e-beam, etc.), detection thresholds, filtering, and / or polarization. it can.

In another embodiment, feedback is used to control the manufacturing process. For example, if too many errors are detected on a particular machine, the feedback from the analyzer can be used to shut down the process or the machine. Similarly, feedback, either static or dynamic, can be used to redirect the lot to the lowest error rate machine or process.

In another embodiment, instead of a separate step, the inspection and analysis / classification step is performed in real time during the inspection step. (An example of this is shown in the system of FIG. 10 (c).) In such a system, the inspection system 1052 is shown inside the system 1004, indicating that it is part of the same system as the analyzer 1056.

FIG. 2 is a block diagram illustrating the interaction of a human user with an embodiment of detection classifier software. A working set of images 208, such as an image with wafer defects, is displayed for user review (as described below in connection with FIG. 3). A human user 210 can review defective images in the working set.

A human user can also request that images be organized by natural grouping 212 and displayed according to this organization. Humans can manually classify defective images into classes (also called "bins") depending on their understanding of the type of defect represented by the image. Currently Kohonen
Mapping techniques are used to naturally group defective images using the features extracted for the defective images. Kohonen mapping is described, for example, in T.W. Kohonen, "The Self-Organizing M
ap ", Proceedings of the IEEE, Vol. 78, 1
990, pp. 1464-1480, which is incorporated herein by reference. Other methods for natural grouping, N.N. Otsu,
"A Threshold Selection Method from G
ray-Level Histograms ", IEEE Trans. Sys
tems, Man, and Cybernetics, Vol. SMC-9, 1
979, pp. 62-66 (incorporated herein by reference), K-means, or any other suitable technique or method for grouping defective images according to common features. Can be used. In the described embodiment, both the natural grouping 212 and the automatic classifier 204 use the same feature set.

Also, a human user can select images from the working set to be placed on the “training set”. Select an image and the user
Manually add images / defects to the set's classes / bins.
Features are extracted from the selected images and stored with the class / bin during the "training classifier" operation. The classifier then classifies the sequence of images (such as set WT) and the user reviews the errors found in the classifier's decision. For example, the user can look at the confusion matrix to determine where the classifier differs from the user's classification. The user can then improve the training set by re-classifying the images, for example by adding, deleting or re-classifying images via a drag-and-drop interface and re-classifying the classifier performance until satisfactory results are achieved. evaluate.

The images in the training set are sent to the feature extractor software 206, which extracts a set of predefined features for each image in the training set. The data structure that stores a series of features for an image is called the "feature vector" for that image. The feature vector contains the value of each feature for a particular image.

The predefined features that are preferably extracted from the training set include, but are not limited to: a) size, brightness, color, shape, texture, moment of inertia, context, proximity to wafer features or to other defects, connectivity to adjacent features or other defects, other yield related images. Characteristics (for example, short,
Features extracted from the image, such as open, bridged, particles, scratched, etc. b) Spatial cluster analysis
A) defect coordinates on the wafer and spatial clusters of defect coordinates, and c) image type information as in lists a) and b), composite or electrical information obtained from analysis techniques, and defects in question. Other information about deductively deployed defects, including, but not limited to, processing history, yield relevance or origin information. It will be appreciated that any suitable feature can be extracted without departing from the spirit of the invention. Examples of analytical techniques used to obtain synthetic or electrical information include a myriad of methods for analyzing defects and a summary of their yield relevance, the “Semiconductor C
harmonization: Present Status & Fut
ure Needs "W. M. Bullis, D. G. Seiler, A .; C
． Diebold, American Institute of Physi
cs 1996, ISBN 1-56396-503-8.

The monitoring automatic dynamic classifier software 204 uses the extracted features of the images in the training set, the working set (W), which was not selected by the user as part of the training set (T). Classify the images in (i.e., classify the set of images WT). In the preferred embodiment of the invention, the classifier 204 uses a nearest neighbor method in which features are extracted from the image set WT and each image in WT is classified as belonging to a class. In general, WT images belong to a class whose components have feature vectors that most closely resemble the image feature vector. It should be appreciated that other automatic classification methods can be used. For example, the features in the feature vector can be weighted (either by the user or with a predefined weighting).

Once the images of set WT have been classified by the classifier 204, the results of automatic classification are compared to the results of manual classification by the user. If the user has classified any images that were also classified by the classifier 204, the results are compared and the comparison is displayed to the user in visual form. At this point, the user can change his classification scheme and make changes to the training set of images accordingly. The user may also change his manual classification if he determines that automatic classification seems more correct.

FIG. 3 illustrates the interface generated by the defect classifier software in accordance with the preferred embodiment of the present invention. In particular, FIG. 3 shows an example of a "Smart Gallery" window according to a preferred embodiment of the present invention. One of the main purposes behind the smart gallery window is to provide gallery-based classification capabilities. smart·
The gallery presents groups of defects in a systematic fashion, allowing the user to view a concise image of defects with adjustable size and resolution. It also assists users in creating classification schemes by providing defects in groups based on their appearance.

The benefits of the smart gallery system include: • Defect sorting • Allowing human users to perform faster and more efficient classification and organization of their manual sorting schemes • Faster manual sorting scheme creation • Shorter manual sorting time • Better Manual classification quality (repeatability) -Faster and more efficient classifier creation process

The window of FIG. 3 includes a toolbar 302 that allows the user to open and save images with classes and defects, classify images with working sets, and create and manipulate training sets. Including commands. Reliability setting area 304
Allows the user to adjust the confidence level, which is an adjustable setting of how close an unknown defect can be to the training set. The value is preferably in the range of 0 (the loosest setting) to 1 (the tightest setting). The changes can be viewed dynamically in the confusion matrix 306.

The confusion matrix 306 is used to display the results of manual versus automatic defect classification. A confusion matrix can be generated for both the current set of images or an explicitly selected set. The results of the manual (human) classification are displayed on the X-axis and the automatic classification by the classifier 204 is displayed on the y-axis. The results of the matching comparisons for all defect classes (if the manual and automatic classification results match) are displayed diagonally on the confusion matrix.

Area 308 displays the working set of images. These images may be displayed in an unsorted order or may be sorted or arranged by natural grouping at the user's option. Users are working set gallery 30
It is preferred to drag and drop images from 8 into the training set gallery 310. Here, the images in the training set are arranged and displayed in a user-specified class. The training set area 312 displays the classes (also referred to as "bins") that contain the composition of the training set, defines groupings, and allows the user to create new classes / bins. When this area 312 is active, the user can use the toolbar 302 to create new classes / bins.

The natural grouping matrix 314 allows the user to see how the images of the working group are distributed in the natural grouping. The number of images in the group is represented by the element number of the matrix 314. The user can click on the elements in the matrix 314 to see all defective images in a particular natural grouping.

In summary, users are classified into working groups classified in natural groupings.
It can optionally indicate that a set of images is desired (eg, via a toolbar or menu item). The user then drags and drops the image from the area 308 into the training set class / bin 310/312 to display the "training" feature. The training function stores the feature vectors of user-selected images of the training set and stores them in association with class / bin. Once the training set is displayed, the automatic classifier classifies the remaining images. The classifier 204 may also use the training set to classify some other set of images. In a different embodiment, the classifier 2
04 can run in the background, reclassify images whenever the training set changes, or be explicitly run by the user.
The classifier 204 results are compared to any manual classifications made by the user, and the comparison results are displayed on the confusion matrix. The user can then add images to the training set or subtract images from the training set via 310 of FIG. 3, depending on the contents of the confusion matrix.

Images are included in the PCT published January 20, 2000, which is hereby incorporated by reference.
Grouping can be done automatically according to immutable core classes, such as those shown in publication WO 00/03234 (inventor: Ben-Porath et al.). The results of the ADC process are incorporated herein by reference, November 1, 1999.
Overall Fab Yield, such as that shown in PCT Publication No. WO 99/59200 (Inventor: Lamey et al.) Published 8th.
) It can also be built into the management system. This application is also integrated by reference to PCT Publication No. WO 99/67626 (inventor: Ravid), which was published on December 29, 1999.

In FIG. 4A, the manual and automatic classifications match (see diagonal element 402).
An example of a confusion matrix 306 is shown. In this example,
One image is matched to class "3", three images are matched to class "2", and one image is matched to class "1". Click the Modify button next to the matrix to highlight matching results (h
light). Clicking on the "Known errors" button will highlight the inconsistent results. Clicking on the "Image" button allows the user to view the image used to generate the particular element of the matrix.

FIG. 4B shows an example of the confusion matrix 306 in which the manual and automatic classifications do not match. Element 45
2 is a non-zero element that is off the diagonal column of the matrix.

FIGS. 5 (a) and 5 (b) show individual examples of interfaces 502 that allow a user to display images in a working set and a training set in a sorted order. , 504 are shown. Images are preferably sorted by factors such as lot number, manual bin, suggested bin, and size.

FIG. 6A is a flowchart showing a method of natural grouping of images in the working set 308. At element 602, the defective image is searched for and the image is captured. Features are extracted from the image at element 604. The extracted features are input to the natural grouping method 606, which can be any suitable method. In the described embodiment, the image feature vectors are grouped using the known Kohonen mapping technique. In the described embodiment, non-random numbers are distributed in the Kohonen map to increase the stability of the grouping and make the grouping repeatable. In some embodiments, the images are displayed in their natural groups (also called clusters), as shown in Figure 6 (b). In other embodiments, the images are arranged to reflect the actual Kohonen map layout.

Particular embodiments are incorporated herein by reference: 1) http: // www
w-ismv. ic. ornl. gov / projects / SSA. html
2) Gleason S.M. S. Tobin K .; W. & Karnowsk
i T. P. , "Spatial Signal Analysis o
f Semiconductor Defects for Manufact
URING PROBLEM DIAGNOSSIS ”, Solid State
Technology, July, 1996, 3) http: // www. d
ym. com / ssa. html, 4) http: // www. electrog
las. com / products / knights_datasheets /
spar_ds. and 5) http: // www. ornl. gov
/ Press_Releases / archive / mr199980804-0
0. html. Spatial signature analysis, as described in
Ignition analysis (SSA) technology is used.

Also, in certain embodiments, analysis and classification is not limited to images, but can be performed on the clusters themselves. In such an embodiment, T.W. P. Karnowski, K .; W. Tobin, S .; S. Gl
eason, Fred Lakhani, SPIE's 24th Annua
l International Symposium on Metrolo
gy, Inspection and Process Control fo
r Microlithography XIII, Santa Clara
Convention Center, Santa Clara, CA, 199
As described in February 1997, the classifier receives "cluster-based features" instead of raw images. Such systems apply grouping and Kohonen mapping to clusters rather than raw images. For non-image data, clustering is collected by EDS specifier (using X-ray system for analysis in e-beam system) or by SSA analysis.

FIG. 7 shows an enlarged view of the training area 312. The user may be in the group displayed in area 310 from working group 308, or in area 312.
Images can be dragged into the class / bin of. In the illustrated embodiment, each class / bin has a class code currently assigned to the class / group, a group code (reflecting its natural group) and some defective images. ing. Of course, the user can add and delete classes / bins as desired (eg via the toolbar).

When the user adds a new class / bin, the class / bin is added. Otherwise, the existing class / bin is opened. The user then manually adds the image / defect to the class. Features are extracted from the selected images and stored during the "classifier training" operation (eg, via the toolbar). The classifier then classifies the sequence of images (such as set WT) and the user reviews the errors found in the classifier's decision. For example, if the user
Looking at the matrix, the classifier can determine what was different from the user's classification. The user can then improve the training set by adding, deleting or reclassifying images, for example via a drag and drop interface, and reassess the classifier performance until satisfactory results are achieved. To do.

FIG. 8 shows a “Smart Gallery” setup function 802 and an automatic classifier production function (auto Classifier Cr).
804) and a classifier function (classifier function)
on) 806, an exemplary user interface is shown. The smart gallery setup feature leads to the user interface of FIG. The classifier function leads to the user interface of FIG. "Smart Gallery" is a trademark of KLA-Tencor Corporation.

FIG. 9 shows an additional user interface for an embodiment of the automatic classifier function. This embodiment is an alternative to the toolbar driven drag and drop method. Using this interface, the user can add defective images to the training set 902 and specify features to extract for natural grouping and for the feature extractor of the classifier 904. The user can specify the number of features to be extracted (here, 80). When the user selects the training button 906, the features of the images in the training set are extracted and saved as a feature vector for each image. The class / bin of each image is
Saved in relation to the feature vector.

The user can set filters 908 on the images to remove specific groups, images, and image types from the feature extraction process. The user can also use button 910 to adjust the confidence of the feature method used by classifier 204.

When the user clicks the test (training set) button 912, the classifier 204 classifies the series of images WT into bins in the training set according to the feature vectors of the images in the training set.

10 (a) and 10 (b) are block diagrams of systems according to the present invention distributed over a network such as the Internet or an intranet.
In FIG. 10 (a), an optical, e-beam, or other type of inspection system 100.
2, the classifiers 1004/104, and the analysis system 1006 (see FIG. 1) are distributed over the network. In FIG. 10 (b), the elements of the classifier 1004, including the natural grouping process 1054, the automatic monitoring classifiers 1056/204, and the feature extractor 1058 are distributed over the network. The natural grouping step 1054 receives the characteristics of the working set as input and outputs the natural grouping of the working set. The auto-monitoring classifier 1056 receives the training set features and classes as well as the defective image features while outputting the defective image classes to be classified. The feature extractor 1058 receives the image and outputs the features of the image.

FIG. 10 (b) also shows an embodiment in which the classifier receives the tool history 1005 as input. The tool history includes, for example, a maintenance history of a tool or machine that performs an inspection process and / or a manufacturing process. If the tool has been repaired according to the proposed maintenance schedule, the data can be weighted more than the data from the unfixed tool. Tool history 1
055 may also include a test value threshold that indicates that the classifier must perform maintenance in order to trust the data from the tool. This threshold may be different for each individual tool and may be the same for all tools of a particular type or function. The tool history can also indicate, for example, whether (or from which tool) two types of semiconductors were obtained from the same tool. Therefore, the tool history 1055 can include the device ID, for example.
For example, if it is known that Tool A had a problem in the past, the data from Tool A can be treated differently than the data from Tool B, which is not a problem.

As mentioned above, FIG. 10 (c) shows that the inspection and analysis / classification steps are performed in real time during the inspection step rather than as separate steps. In such a system, inspection system 1052 is shown inside system 1004 to show that it is part of the same system as classifier 1056.
This system utilizes review images (“patch” images) to achieve real-time inspection. Newer inspection systems, such as the KLA2350, include ADCs embedded in an image computer that operate in "real time" during inspection. iA
This system, called the DC (Integrated ADC) system, works by grabbing "patches" around the defect location during the scan and performing an ADC on the defective pixels in these patches. All this is done in hardware within the tester so that no additional throughput is required.

FIGS. 10 (a), 10 (b), and 10 (c) show that in a particular system,
The classifier includes dotted lines 1003, 1053, respectively, which indicate that a feedback signal can be provided to the inspection system in a manner similar to that described above in connection with FIG.

It will be appreciated that various embodiments and modifications may be made without departing from the spirit and scope of the invention. For example, the inventive concept automatically classifies images in the background by defect type (eg, while running defect analysis software), and then distributes each selected type throughout. It can be expanded to include displaying the results as the wafer map shown. This is just one way to use the output data. Since the defect location distribution can be useful in identifying defect sources, the ability to select similar defects combined with its ability to see its spatial distribution (natural grouping) can be powerful. K
For each cluster of the Ohen map, a display can be provided that shows a) a representative image and b) a defect map showing the location of defects in the cluster.

From the above description, it will be apparent that the invention disclosed herein provides a new and advantageous system and method for optical inspection used to classify semiconductor defects.

[Brief description of drawings]

1a is a schematic block diagram of a semiconductor optics, e-beam, or other type of inspection system. FIG. FIG. 1b is a schematic block diagram of a semiconductor optics, e-beam, or other type of inspection system that employs the present invention for inspection setup.

FIG. 2 is a block diagram illustrating the interaction of a human user with a section of an embodiment of the defect classifier software.

FIG. 3 is a diagram showing an interface generated by defect classifier software according to a preferred embodiment of the present invention.

FIG. 4a is a diagram showing an example of a confusion matrix in a user interface with matching manual and automatic classification. FIG. 4b is a diagram showing an example of a confusion matrix in a user interface where manual and automatic classification do not match.

FIG. 5a is a diagram showing an example interface that allows a user to display images in a sorted order in a working set and a training set. FIG. 5b is a diagram illustrating an example interface that allows a user to view images in a sorted order in a working set and a training set.

FIG. 6a is a flow chart showing a method of natural grouping of images in a working set. FIG. 6b shows an image organized and displayed in natural groups.

FIG. 7 shows an enlarged view of the training area in the user interface.

FIG. 8 illustrates an exemplary user interface used in the preferred embodiment of the present invention.

[Figure 9]   FIG. 7 illustrates an additional user interface for classifier functionality.

FIG. 10a is a block diagram of a system according to the present invention distributed over a network such as the Internet or an intranet. FIG. 10b is a block diagram of a system according to the present invention distributed over a network such as the Internet or an intranet. FIG. 10c is a block diagram of a system according to the present invention distributed over a network such as the Internet or an intranet.

─────────────────────────────────────────────────── ─── Continued front page    (81) Designated countries EP (AT, BE, CH, CY, DE, DK, ES, FI, FR, GB, GR, IE, I T, LU, MC, NL, PT, SE, TR), OA (BF , BJ, CF, CG, CI, CM, GA, GN, GW, ML, MR, NE, SN, TD, TG), AP (GH, G M, KE, LS, MW, MZ, SD, SL, SZ, TZ , UG, ZW), EA (AM, AZ, BY, KG, KZ, MD, RU, TJ, TM), AE, AG, AL, AM, AT, AU, AZ, BA, BB, BG, BR, BY, B Z, CA, CH, CN, CR, CU, CZ, DE, DK , DM, DZ, EE, ES, FI, GB, GD, GE, GH, GM, HR, HU, ID, IL, IN, IS, J P, KE, KG, KP, KR, KZ, LC, LK, LR , LS, LT, LU, LV, MA, MD, MG, MK, MN, MW, MX, MZ, NO, NZ, PL, PT, R O, RU, SD, SE, SG, SI, SK, SL, TJ , TM, TR, TT, TZ, UA, UG, UZ, VN, YU, ZA, ZW (72) Inventor Banazy, Cybal             California 94536,             Fremont, Litchfield Aveni             View 1250 (72) Inventor Smith, Ian             United States California 95032,             Los Gatos, Dover Street             228 F-term (reference) 4M106 CA38 DJ40

Claims (40)

[Claims] 1. A method implemented by a data processing system for classifying a plurality of received images, the method comprising extracting features from a training set that is a user-selected subset of the plurality of images. Each image of the set has a concatenated class; the data processing system classifies at least one of the plurality of images according to the extracted features and classes of the training set; Of one of the images of, and displaying the results of the comparison between the classification by the data processing system and the classification by the user. 2. The method of claim 1, wherein the features of the training set include size. 3. The method of claim 1, wherein the features of the training set include brightness. 4. The feature of the training set comprises color.
The method described in. 5. The method of claim 1, wherein the features of the training set include shapes. 6. The features of the training set include textures, moments of inertia, contexts, proximity to wafer features, proximity to other defects,
The method of claim 1, comprising at least one of connectivity to adjacent features, connectivity to other defects, and yield-related properties obtained from the image. 7. The method of claim 1, wherein the features of the training set include defect coordinates on a wafer. 8. The method of claim 1, wherein the features of the training set include defect coordinates when spatial cluster analysis is used. 9. The method of claim 1, wherein the features of the training set include information obtained from one of processing history, yield, relevance, and origin of defects. 10. Classifying at least one of the plurality of images by the data processing system according to extracted features and classes of the training set includes classifying the plurality of images using Kohonen map technology. The method of claim 1, comprising: 11. The method according to claim 10, wherein a non-random number is distributed in the Kohonen map. 12. Categorizing at least one of the plurality of images by the data processing system according to extracted features and classes of the training set includes using the spatial signature analysis technique to classify the plurality of images. The method of claim 1, comprising classifying. 13. The method of claim 1, wherein classifying at least one of the plurality of images by the data processing system further comprises classifying according to cluster-based features instead of images. 14. Allowing a user to classify one of the plurality of images includes displaying the images to the user in a classification group determined by the classifying step. The method of claim 1. 15. The method of claim 1, further comprising sending feedback to the inspection system to fine tune the inspection system according to user classification. 16. The inspection object is inspected in real time, and the result of the inspection set is
The method of claim 1, further comprising sending to a classifier trained according to the plurality of images classified by a user. 17. The method of claim 1, wherein the features include tool history information associated with an inspection system. 18. The method of claim 1, wherein the features include tool history information related to past success rates for the classifying step. 19. The method of claim 1, wherein only some of the plurality of images are associated with a semiconductor etching process. 20. A method implemented by a data processing system for classifying a plurality of received images, the method comprising extracting features from a training set that is a user-selected subset of the plurality of images. Each image of the set has an associated class, the data processing system classifies at least one of the plurality of images according to the extracted features and classes of the training set, by the data processing system Displaying the result of the comparison between the classification and the classification by the user. 21. A system for classifying a plurality of received images, the software portion configured to extract features from a training set that is a user-selected subset of the plurality of images. Each image of a software part having an associated class, and a software part configured by the system to classify at least one of the plurality of images according to the extracted features and classes of the training set. A software portion configured to allow a user to classify one of the plurality of images; and to send feedback to the inspection system to fine tune the inspection system according to the user's classification. A system that includes a configured software portion. 22. The system of claim 21, wherein the features of the training set include size. 23. The system of claim 21, wherein the features of the training set include brightness. 24. The system of claim 21, wherein the features of the training set include color. 25. The system of claim 21, wherein the features of the training set include shapes. 26. The features of the training set are textures, moments of inertia, contexts, proximity to wafer features, proximity to other defects, connectivity to adjacent features, connectivity to other defects. , And at least one of a yield-related characteristic obtained from the image. 27. The system of claim 21, wherein the features of the training set include defect coordinates on a wafer. 28. The system of claim 21, wherein the features of the training set include defect coordinates when spatial cluster analysis is used. 29. The features of the training set include information obtained from one of processing history, yield, relevance, and source of defects.
The system according to 1. 30. The portion configured to classify at least one of the plurality of images by the data processing system according to extracted features and classes of the training set using the Kohonen map technique. 22. The system of claim 21, including a portion configured to classify the plurality of images. 31. The system of claim 30, wherein the Kohonen map is distributed with non-random numbers. 32. A portion configured to classify at least one of the plurality of images by the data processing system according to extracted features and classes of the training set using spatial signature analysis techniques. 22. The system of claim 21, including a portion configured to classify the plurality of images. 33. The portion configured to classify at least one of the plurality of images by the data processing system further comprises a portion configured to classify according to a cluster-based feature instead of the image. 22. The system of claim 21, comprising. 34. The portion configured to allow a user to classify one of the plurality of images displays the image to the user in the classification group determined by the classifying step. 22. The system of claim 21, including a portion configured to: 35. The system of claim 21, further comprising a portion configured to send feedback to the inspection system to fine tune the inspection system according to user classification. 36. The inspection target is inspected in real time, and the result of the inspection set is
22. The system of claim 21, further comprising a portion configured to send to a classifier trained according to the plurality of images classified by a user. 37. The system of claim 21, wherein the features include tool history information associated with an inspection system. 38. The system of claim 21, wherein the features include tool history information related to past success rates for the classifying step. 39. The system of claim 21, wherein only some of the plurality of images are associated with a semiconductor etching process. 40. A system for classifying a plurality of received images, the software portion being configured to extract features from a training set that is a user-selected subset of the plurality of images. A portion of each of the images having an associated class, and a portion configured by the system to classify at least one of the plurality of images according to the extracted features and classes of the training set. To the inspection system to fine tune the inspection system according to the classification made by the data processing system.